Here's a patch (force_to_memory2.patch) that I'm hoping fixes the cmath 
test failures on Solaris 10/x86.

Skip, could you give it a try?

The patch isn't final:  I need to look for more places where 
Py_FORCE_DOUBLE should be applied.  But I'll wait to find out whether this 
really does fix the problem, first;  I'm still just guessing about the 
cause.

By the way, it looks like the problem with the original patch, on OS X, 
was that nothing in pymath.c is used in the Python *core*, so the symbols 
from pymath.o aren't compiled into the python.exe executable, and they're 
not available when loading modules.  Rather than working out how to fix 
this, I just moved the definitions into Objects/floatobject.c instead.